Robot Foundationism
From the AI Research Compass wiki
Robot Foundationism: Believes general robot skill will come from foundation models trained on large, diverse real-world robot data, gathered through a deployment flywheel.
Beliefs
Robot Foundationism applies the foundation-model recipe to physical action, with one stubborn amendment: the data must come from real robots in the real world. Language models had the internet; robots have nothing comparable, so the data must be manufactured through teleoperation, fleets of deployed machines and pooled datasets across labs and embodiments. Train a large vision-language-action policy on all of it, deploy, collect more, retrain. The flywheel, not a clever algorithm, is what will produce generalist robots.
It sits near the far World edge and close to the Monolith end: a single generalist policy, though one embedded in an operational system of data collection and deployment. Its paradigm level is Steady. It treats robotics as a scaling problem, a break from decades of hand-engineered controllers but in line with today’s mainstream recipe, and it is pragmatic about the messy logistics that scaling requires.
Arguments
- Transformer policies trained on large sets of real demonstrations show improved generalization to new objects, scenes and instructions as data diversity grows.
- Pooling data across many robot embodiments yields positive transfer, so each lab’s data helps everyone’s robot.
- Real-world data avoids the sim-to-real gap entirely; what the policy learns is what the robot will actually face.
Critiques
- Real data is slow and costly; even large robot datasets are tiny next to web text, and scaling may stall.
- Sim-to-real camps argue simulation offers orders of magnitude more experience at lower cost and with safe failure.
- Imitation from demonstrations caps performance at the demonstrator’s skill and struggles with recovery from novel errors.
Representative works
- RT-1: Robotics Transformer for Real-World Control at Scale2022
- Open X-Embodiment: Robotic Learning Datasets and RT-X Models2023
- π0: A Vision-Language-Action Flow Model for General Robot Control2024
Relationships
Friends
Frenemies
How to get it
Land on this cell in the full or pro test: x ≈ +0.99, y ≈ +0.74, paradigm ≈ −0.36.